arxiv: 2605.13032 · v2 · submitted 2026-05-13 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition

Danny Wang , Ruihong Qiu , Zi Huang

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:37 UTC · model grok-4.3

classification 💻 cs.LG

keywords graph neural networksout-of-distribution detectioninformation decompositioninformation bottlenecknode classificationgraph structurefeature informationspurious correlations

0 comments

The pith

TIDE decomposes graph information into feature-specific, structure-specific and joint components to retain only label-relevant joint signals for improved out-of-distribution node detection.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces TIDE to address vulnerabilities in graph neural networks for node classification when facing out-of-distribution shifts in features or structure. Standard supervised learning captures spurious signals, but TIDE breaks the information down into three components: feature-specific, structure-specific, and joint. It keeps the part of the joint information tied to true labels and removes the spurious unique-to-feature or unique-to-structure parts. This creates a clearer distinction between in-distribution and out-of-distribution nodes, with higher confidence on ID data and larger entropy differences. Readers should care because it shows a way to make graph models more robust by explicitly managing what information is used rather than letting the model pick up misleading patterns.

Core claim

TIDE explicitly decomposes information into feature-specific, structure-specific and joint components, preserving only the label-relevant part of the joint information while filtering out spurious feature- and structure-specific information, thereby enhancing the separation between in-distribution (ID) and OOD nodes. Beyond the framework, theoretical and empirical analyses show that an information bottleneck objective is preferable to standard SL for graph OOD detection, with higher ID confidence and a greater entropy gap between ID and OOD data.

What carries the argument

The Tri-Component Information Decomposition framework that separates node information into feature-specific, structure-specific, and joint components and applies an information bottleneck to retain only label-relevant joint information.

If this is right

An information bottleneck objective produces higher ID confidence and a larger entropy gap between ID and OOD data than standard supervised learning.
TIDE improves FPR95 by up to 34% over strong baselines across seven datasets without sacrificing ID accuracy.
Filtering out spurious feature- and structure-specific information leads to better ID-OOD separation in graph node classification.
The approach reduces vulnerability to distributional changes in node features and graph structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This decomposition strategy could apply to other graph learning tasks where spurious correlations between features and labels need explicit removal.
If the tri-component separation proves stable, it opens questions about whether similar information accounting applies to non-graph data like sequences or images.
Testing on graphs with known causal structures would verify if the joint component truly captures label-relevant information.

Load-bearing premise

The joint information component can be reliably separated into label-relevant versus spurious parts such that removing the spurious specific components creates a measurable entropy gap and higher ID confidence without reducing ID accuracy.

What would settle it

Observing that TIDE does not increase the entropy gap between ID and OOD nodes or that it lowers ID classification accuracy on standard benchmarks would disprove the main claim.

Figures

Figures reproduced from arXiv: 2605.13032 by Danny Wang, Ruihong Qiu, Zi Huang.

**Figure 1.** Figure 1: Comparison of information captured by standard supervised learning (SL), GIB (Wu et al., 2020), and TIDE. The dashed red circle shows that SL mixes label-relevant and irrelevant feature and structure signals, causing spurious correlations that make OOD nodes appear ID. The gridded region indicates that GIB suppresses label-irrelevant noise but still keeps input-specific spurious cues. Contrary, TIDE (solid… view at source ↗

**Figure 2.** Figure 2: ID-OOD representation distributions for standard SL trained detector and TIDE’s joint classification (Cls.), structure (Struc.) and feature (Feat.) networks on Cora-Structure with a structure shift and ID-like features (XID, AOOD). ducing a compact and shift-indicative representation. We provide theoretical insights showing that, compared to SL, IB increases ID confidence and improves ID-OOD separation fo… view at source ↗

**Figure 3.** Figure 3: Framework of TIDE: an information decomposition approach that preserves joint label-relevant information in Z while suppressing spurious feature-only (V) and structure-only (Q) signals. By Eq. 6, the mutual information I(X, A; Y) thus becomes: I(X, A; Y) = I(Z; Y) | {z } joint + I(X; Y|Z) | {z } feature + I(A; Y|Z) | {z } structural . (8) This illustrates our tri-component information decomposition paradi… view at source ↗

**Figure 4.** Figure 4: Prediction confidence and energy score distributions of GNNSAFE and TIDE on Cora-Feature; dashed line shows FPR95. Additionally, the second and third subplot shows that TIDE achieves a greater separation between ID and OOD energy scores than GNNSAFE. This increased energy margin highlights TIDE’s efficacy in distinguishing ID from OOD data. The dashed FPR95 thresholds indicate reduced ID-OOD overlap for … view at source ↗

read the original abstract

Graph neural networks are widely used for node classification, but they remain vulnerable to out-of-distribution (OOD) shifts in node features and graph structure. Prior work established that methods trained with standard supervised learning (SL) objectives tend to capture spurious signals from either features and/or structure, leaving the model fragile under distributional changes. To address this, we propose TIDE, a novel and effective Tri-Component Information Decomposition framework that explicitly decomposes information into feature-specific, structure-specific and joint components. TIDE aims to preserve only the label-relevant part of the joint information while filtering out spurious feature- and structure-specific information, thereby enhancing the separation between in-distribution (ID) and OOD nodes. Beyond the framework, we provide theoretical and empirical analyses showing that an information bottleneck objective is preferable to standard SL for graph OOD detection, with higher ID confidence and a greater entropy gap between ID and OOD data. Extensive experiments across seven datasets confirm the efficacy of TIDE, achieving up to a 34% improvement in FPR95 over strong baselines while maintaining competitive ID accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TIDE's tri-component decomposition for graph OOD detection is a fresh framing but rests on an unproven identifiability step that the abstract does not resolve.

read the letter

The core idea is to split node information into feature-specific, structure-specific, and joint terms, then apply an information bottleneck to keep only the label-relevant slice of the joint while discarding the rest. This targets the known weakness of standard supervised GNN training picking up spurious signals under feature or structure shifts. The paper shows this produces a larger entropy gap between ID and OOD nodes and reports up to 34% better FPR95 on seven datasets while holding ID accuracy steady. That empirical pattern is the strongest part of the work and gives a concrete handle on the robustness problem. The theoretical claim that the bottleneck is preferable to plain supervised learning for this task is also worth testing. The soft spot is identifiability. Nothing in the provided material shows that the three components are uniquely recoverable from the observed joint distribution or that the bottleneck cleanly isolates the label-relevant joint part without extra assumptions on how the shift occurs. If the decomposition depends on quantities fitted from the same labels it aims to generalize beyond, the circularity risk noted in the stress-test note remains live. The paper would benefit from a clear derivation or counter-example showing when the separation succeeds or fails. Readers working on graph robustness or information-theoretic regularization will find the framework useful to try, even if the theory needs tightening. It is worth sending to peer review because the problem is practical, the experiments are broad, and the proposed fix is specific enough to critique directly.

Referee Report

2 major / 2 minor

Summary. The paper proposes TIDE, a Tri-Component Information Decomposition framework for graph out-of-distribution detection in node classification. It decomposes node information into feature-specific, structure-specific, and joint components, then applies an information-bottleneck objective to retain only the label-relevant slice of the joint term while discarding spurious feature- and structure-specific information. Theoretical analyses argue that the IB objective yields higher ID confidence and a larger entropy gap than standard supervised learning, and experiments across seven datasets report up to 34% FPR95 improvement while preserving competitive ID accuracy.

Significance. If the decomposition is identifiable and the filtering step demonstrably preserves label-predictive signal without introducing new assumptions on the shift, the work would supply a principled mechanism for mitigating spurious correlations in both node features and graph structure. The explicit comparison of IB versus SL objectives and the reported empirical gains could influence subsequent graph OOD methods that seek to control information flow rather than rely solely on post-hoc scoring.

major comments (2)

[§3] §3 (Method): The tri-component decomposition is asserted to separate feature-specific, structure-specific, and joint information from p(X, A, Y), yet no derivation establishes unique recoverability of the three terms. Without identifiability, the subsequent claim that discarding the two specific components leaves only label-relevant joint information cannot be guaranteed, directly affecting the promised entropy gap and OOD gains.
[§4] §4 (Theoretical Analysis): The information-bottleneck Lagrangian is said to isolate the label-relevant portion of the joint component, but the text provides no explicit conditions (e.g., conditional independence of spurious factors from Y) under which this separation holds. This assumption is load-bearing for the assertion that IB is preferable to standard supervised learning for OOD detection.

minor comments (2)

[Abstract] Abstract: The maximum 34% FPR95 improvement is stated without naming the strongest baseline or the dataset on which it occurs; adding this detail would improve reproducibility of the headline result.
[§3] Notation: The symbols used for the three information components (feature-specific, structure-specific, joint) are introduced without an accompanying table that maps them to the corresponding mutual-information expressions; a compact notation table would aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on the theoretical foundations of TIDE. We address each major comment below and will revise the manuscript accordingly to strengthen the derivations and assumptions.

read point-by-point responses

Referee: [§3] §3 (Method): The tri-component decomposition is asserted to separate feature-specific, structure-specific, and joint information from p(X, A, Y), yet no derivation establishes unique recoverability of the three terms. Without identifiability, the subsequent claim that discarding the two specific components leaves only label-relevant joint information cannot be guaranteed, directly affecting the promised entropy gap and OOD gains.

Authors: We acknowledge that the current manuscript does not include an explicit derivation establishing unique recoverability of the three components. The decomposition is defined via the partial information decomposition (PID) of I(X,A;Y) into unique feature-specific, unique structure-specific, and joint (redundant/synergistic) terms. We will add a new subsection in §3 that derives these quantities directly from the joint p(X,A,Y) using the standard PID lattice and shows how the subsequent IB objective filters the joint term. While exact identifiability in finite samples may require additional regularity conditions on the encoders, the operational procedure (minimizing I(spurious;Z) while maximizing I(joint;Y)) remains well-defined and is supported by the empirical results across seven datasets. revision: yes
Referee: [§4] §4 (Theoretical Analysis): The information-bottleneck Lagrangian is said to isolate the label-relevant portion of the joint component, but the text provides no explicit conditions (e.g., conditional independence of spurious factors from Y) under which this separation holds. This assumption is load-bearing for the assertion that IB is preferable to standard supervised learning for OOD detection.

Authors: We agree that the manuscript should state the conditions under which the IB objective isolates the label-relevant joint information. In the revision we will add a formal statement (new Theorem in §4) that assumes (i) the spurious feature and structure components are conditionally independent of Y given the joint component, and (ii) the encoders can approximate the relevant mutual-information terms. Under these conditions we prove that the IB Lagrangian yields strictly higher ID confidence and a larger entropy gap than standard supervised learning. We will also include a brief discussion of robustness when the conditional-independence assumption is mildly violated, consistent with the observed empirical gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The abstract presents TIDE as a proposed framework that decomposes node information into three components and applies an information-bottleneck objective to retain only the label-relevant joint slice. No equations, self-citations, or fitted-parameter renamings are exhibited that reduce the claimed entropy gap or OOD improvement to a tautology or to quantities defined by the same supervised objective. The theoretical analysis favoring IB over SL is stated as independent content, and the empirical results on seven datasets are external to the decomposition definition. The derivation therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the framework itself is the proposed contribution rather than a new physical entity.

pith-pipeline@v0.9.0 · 5487 in / 1131 out tokens · 37721 ms · 2026-05-15T05:37:48.342808+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

TIDE explicitly decomposes information into feature-specific, structure-specific and joint components... IBZ = max I(Z;Y) − βZ I(X,A;Z)
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

I(X,A;Y) = I(Z;Y) + I(X;Y|Z) + I(A;Y|Z) with A ⊥⊥ X | Z

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

133 extracted references · 133 canonical work pages

[1]

Your classifier is secretly an energy based model and you should treat it like one , booktitle =

Will Grathwohl and Kuan. Your classifier is secretly an energy based model and you should treat it like one , booktitle =

work page
[2]

Srikant , title =

Shiyu Liang and Yixuan Li and R. Srikant , title =. ICLR , year =

work page
[3]

CVPR , year =

Ziqian Lin and Sreya Dutta Roy and Yixuan Li , title =. CVPR , year =

work page
[4]

NeurIPS , year =

Haoran Wang and Weitang Liu and Alex Bocchieri and Yixuan Li , title =. NeurIPS , year =

work page
[5]

NeurIPS , year =

Yiyou Sun and Chuan Guo and Yixuan Li , title =. NeurIPS , year =

work page
[6]

Neural Mean Discrepancy for Efficient Out-of-Distribution Detection , booktitle =

Xin Dong and Junfeng Guo and Ang Li and Wei. Neural Mean Discrepancy for Efficient Out-of-Distribution Detection , booktitle =

work page
[7]

ECCV , year =

Yiyou Sun and Yixuan Li , title =. ECCV , year =

work page
[8]

ICLR , year =

Andrija Djurisic and Nebojsa Bozanic and Arjun Ashok and Rosanne Liu , title =. ICLR , year =

work page
[9]

Owens and Yixuan Li , title =

Weitang Liu and Xiaoyun Wang and John D. Owens and Yixuan Li , title =. NeurIPS , year =

work page
[10]

NeurIPS , year =

Sangha Park and Jisoo Mok and Dahuin Jung and Saehyung Lee and Sungroh Yoon , title =. NeurIPS , year =

work page
[11]

NeurIPS , year =

Jianing Zhu and Yu Geng and Jiangchao Yao and Tongliang Liu and Gang Niu and Masashi Sugiyama and Bo Han , title =. NeurIPS , year =

work page
[12]

CoRR , year =

Gleb Bazhenov and Sergei Ivanov and Maxim Panov and Alexey Zaytsev and Evgeny Burnaev , title =. CoRR , year =

work page
[13]

Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classification , booktitle =

Maximilian Stadler and Bertrand Charpentier and Simon Geisler and Daniel Z. Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classification , booktitle =

work page
[14]

Uncertainty Aware Semi-Supervised Learning on Graph Data , booktitle =

Xujiang Zhao and Feng Chen and Shu Hu and Jin. Uncertainty Aware Semi-Supervised Learning on Graph Data , booktitle =

work page
[15]

WSDM , year =

Yixin Liu and Kaize Ding and Huan Liu and Shirui Pan , title =. WSDM , year =

work page
[16]

KDD , year =

Yuxin Guo and Cheng Yang and Yuluo Chen and Jixi Liu and Chuan Shi and Junping Du , title =. KDD , year =

work page
[17]

ICLR , year =

Qitian Wu and Yiting Chen and Chenxiao Yang and Junchi Yan , title =. ICLR , year =

work page
[18]

NeurIPS , year =

Zenan Li and Qitian Wu and Fan Nie and Junchi Yan , title =. NeurIPS , year =

work page
[19]

ICCV , year =

Aming Wu and Da Chen and Cheng Deng , title =. ICCV , year =

work page
[20]

NeurIPS , year =

Xuefeng Du and Yiyou Sun and Jerry Zhu and Yixuan Li , title =. NeurIPS , year =

work page
[21]

ICML , year =

Seul Lee and Jaehyeong Jo and Sung Ju Hwang , title =. ICML , year =

work page
[22]

DiGress: Discrete Denoising diffusion for graph generation , booktitle =

Cl. DiGress: Discrete Denoising diffusion for graph generation , booktitle =

work page
[23]

Haoyang Li and Xin Wang and Ziwei Zhang and Wenwu Zhu , title =

work page
[24]

Kipf and Max Welling , title =

Thomas N. Kipf and Max Welling , title =. ICLR , year =

work page
[25]

CVPR , year =

Junchi Yu and Jian Liang and Ran He , title =. CVPR , year =

work page
[26]

Hamilton and Zhitao Ying and Jure Leskovec , title =

William L. Hamilton and Zhitao Ying and Jure Leskovec , title =. NeurIPS , year =

work page
[27]

NeurIPS , year =

Shurui Gui and Xiner Li and Limei Wang and Shuiwang Ji , title =. NeurIPS , year =

work page
[28]

NeurIPS Workshop , year=

Mucong Ding and Kezhi Kong and Jiuhai Chen and John Kirchenbauer and Micah Goldblum and David Wipf and Furong Huang and Tom Goldstein , title=. NeurIPS Workshop , year=

work page
[29]

DrugOOD: Out-of-Distribution

Yuanfeng Ji and Lu Zhang and Jiaxiang Wu and Bingzhe Wu and Long. DrugOOD: Out-of-Distribution. CoRR , year =

work page
[30]

NeurIPS , year =

Yongqiang Chen and Yonggang Zhang and Yatao Bian and Han Yang and Kaili Ma and Binghui Xie and Tongliang Liu and Bo Han and James Cheng , title =. NeurIPS , year =

work page
[31]

NeurIPS , year =

Yangze Zhou and Gitta Kutyniok and Bruno Ribeiro , title =. NeurIPS , year =

work page
[32]

CoRR , year =

Haoyang Li and Xin Wang and Ziwei Zhang and Wenwu Zhu , title =. CoRR , year =

work page
[33]

NeurIPS , year =

Nianzu Yang and Kaipeng Zeng and Qitian Wu and Xiaosong Jia and Junchi Yan , title =. NeurIPS , year =

work page
[34]

Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future , journal =

David Ahmedt. Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future , journal =

work page
[35]

A Benchmark of Medical Out of Distribution Detection , journal =

Tianshi Cao and Chinwei Huang and David Yu. A Benchmark of Medical Out of Distribution Detection , journal =

work page
[36]

Patterns , volume =

Seungyeon Lee and Changchang Yin and Ping Zhang , title =. Patterns , volume =

work page
[37]

Harnessing the power of synthetic data in healthcare: innovation, application, and privacy , journal =

Mauro Giuffr. Harnessing the power of synthetic data in healthcare: innovation, application, and privacy , journal =

work page
[38]

CoRR , year =

Jingkang Yang and Kaiyang Zhou and Yixuan Li and Ziwei Liu , title =. CoRR , year =

work page
[39]

Liu and Emily Fertig and Jasper Snoek and Ryan Poplin and Mark A

Jie Ren and Peter J. Liu and Emily Fertig and Jasper Snoek and Ryan Poplin and Mark A. DePristo and Joshua V. Dillon and Balaji Lakshminarayanan , title =. NeurIPS , year =

work page
[40]

CoRR , year =

Hao Lang and Yinhe Zheng and Yixuan Li and Jian Sun and Fei Huang and Yongbin Li , title =. CoRR , year =

work page
[41]

NeurIPS , year =

Julian Bitterwolf and Alexander Meinke and Matthias Hein , title =. NeurIPS , year =

work page
[42]

Novelty Detection Via Blurring , booktitle =

Sung. Novelty Detection Via Blurring , booktitle =

work page
[43]

ECML PKDD , year =

Jiefeng Chen and Yixuan Li and Xi Wu and Yingyu Liang and Somesh Jha , title =. ECML PKDD , year =

work page
[44]

Generalized

Yen. Generalized. CVPR , year =

work page
[45]

CVPR , year =

Matthias Hein and Maksym Andriushchenko and Julian Bitterwolf , title =. CVPR , year =

work page
[46]

Willke , title =

Apoorv Vyas and Nataraj Jammalamadaka and Xia Zhu and Dipankar Das and Bharat Kaul and Theodore L. Willke , title =. ECCV , year =

work page
[47]

ICLR , year =

Kimin Lee and Honglak Lee and Kibok Lee and Jinwoo Shin , title =. ICLR , year =

work page
[48]

Hybrid Energy Based Model in the Feature Space for Out-of-Distribution Detection , booktitle =

Marc Lafon and Elias Ramzi and Cl. Hybrid Energy Based Model in the Feature Space for Out-of-Distribution Detection , booktitle =

work page
[49]

ICML , year =

Xue Jiang and Feng Liu and Zhen Fang and Hong Chen and Tongliang Liu and Feng Zheng and Bo Han , title =. ICML , year =

work page
[50]

NeurIPS , year =

Wenjian Huang and Hao Wang and Jiahao Xia and Chengyan Wang and Jianguo Zhang , title =. NeurIPS , year =

work page
[51]

ICLR , year =

Dan Hendrycks and Kevin Gimpel , title =. ICLR , year =

work page
[52]

NeurIPS , year =

Kimin Lee and Kibok Lee and Honglak Lee and Jinwoo Shin , title =. NeurIPS , year =

work page
[53]

CoRR , year =

Zhihao Ding and Jieming Shi , title =. CoRR , year =

work page
[54]

KDD , year =

Yu Song and Donglin Wang , title =. KDD , year =

work page
[55]

ICLR , year=

Longfei Ma and Yiyou Sun and Kaize Ding and Fei Wu , title=. ICLR , year=

work page
[56]

AAAI , year =

Luzhi Wang and Dongxiao He and He Zhang and Yixin Liu and Wenjie Wang and Shirui Pan and Di Jin and Tat. AAAI , year =

work page
[57]

IJCAI , year =

Tiancheng Huang and Donglin Wang and Yuan Fang and Zhengyu Chen , title =. IJCAI , year =

work page
[58]

LoG , year =

Tianjin Huang and Tianlong Chen and Meng Fang and Vlado Menkovski and Jiaxu Zhao and Lu Yin and Yulong Pei and Decebal Constantin Mocanu and Zhangyang Wang and Mykola Pechenizkiy and Shiwei Liu , title =. LoG , year =

work page
[59]

Thiagarajan , title =

Puja Trivedi and Mark Heimann and Rushil Anirudh and Danai Koutra and Jayaraman J. Thiagarajan , title =. ICLR , year =

work page
[60]

ICDM , year =

Lina Yang and Bin Lu and Xiaoying Gan , title =. ICDM , year =

work page
[61]

ICLR , year=

Kuan Li and YiWen Chen and Yang Liu and Jin Wang and Qing He and Minhao Cheng and Xiang Ao , title=. ICLR , year=

work page
[62]

CoRR , year=

Han Wang and Yixuan Li , title=. CoRR , year=

work page
[63]

CoRR , year =

Xu Shen and Yili Wang and Kaixiong Zhou and Shirui Pan and Xin Wang , title =. CoRR , year =

work page
[64]

Dietterich , title =

Dan Hendrycks and Mantas Mazeika and Thomas G. Dietterich , title =. ICLR , year =

work page
[65]

NeurIPS , year =

Haotian Zheng and Qizhou Wang and Zhen Fang and Xiaobo Xia and Feng Liu and Tongliang Liu and Bo Han , title =. NeurIPS , year =

work page
[66]

ICLR , year =

Xuefeng Du and Zhen Fang and Ilias Diakonikolas and Yixuan Li , title =. ICLR , year =

work page
[67]

Outlier Exposure with Focal Loss for Out-of-distribution Detection , booktitle =

Qichao Chen and Zhiyuan Chen and Tom. Outlier Exposure with Focal Loss for Out-of-distribution Detection , booktitle =

work page
[68]

Neurocomputing , volume =

Jiin Koo and Sungjoon Choi and Sangheum Hwang , title =. Neurocomputing , volume =

work page
[69]

Outlier exposure with confidence control for out-of-distribution detection , volume =

Aristotelis. Outlier exposure with confidence control for out-of-distribution detection , volume =. Neurocomputing , year =

work page
[70]

ICLR , year =

Xuefeng Du and Zhaoning Wang and Mu Cai and Yixuan Li , title =. ICLR , year =

work page
[71]

ICLR , year =

Leitian Tao and Xuefeng Du and Jerry Zhu and Yixuan Li , title =. ICLR , year =

work page
[72]

CoRR , year =

Sachin Vernekar and Ashish Gaurav and Vahdat Abdelzad and Taylor Denouden and Rick Salay and Krzysztof Czarnecki , title =. CoRR , year =

work page
[73]

Input Complexity and Out-of-distribution Detection with Likelihood-based Generative Models , booktitle =

Joan Serr. Input Complexity and Out-of-distribution Detection with Likelihood-based Generative Models , booktitle =

work page
[74]

NeurIPS , year =

Zhisheng Xiao and Qing Yan and Yali Amit , title =. NeurIPS , year =

work page
[75]

Wipf and Jun Zhu , title =

Ziyu Wang and Bin Dai and David P. Wipf and Jun Zhu , title =. NeurIPS , year =

work page
[76]

Nalisnick and Akihiro Matsukawa and Yee Whye Teh and Dilan G

Eric T. Nalisnick and Akihiro Matsukawa and Yee Whye Teh and Dilan G. Do Deep Generative Models Know What They Don't Know? , booktitle =

work page
[77]

NeurIPS , year =

Robin Schirrmeister and Yuxuan Zhou and Tonio Ball and Dan Zhang , title =. NeurIPS , year =

work page
[78]

CoRR , year =

Iakovos Evdaimon and Giannis Nikolentzos and Michail Chatzianastasis and Hadi Abdine and Michalis Vazirgiannis , title =. CoRR , year =

work page
[79]

CoRR , year =

Cai Zhou and Xiyuan Wang and Muhan Zhang , title =. CoRR , year =

work page
[80]

NeurIPS , year =

Jonathan Ho and Ajay Jain and Pieter Abbeel , title =. NeurIPS , year =

work page

Showing first 80 references.